Rationality leads people to imitate those with similar tastes but different information. But people who imitate common sources develop correlated beliefs, and rationality demands that later social learners take this redundancy into account. This implies severe limits to rational imitation. We show that (i) in most natural observation structures besides the canonical single-file case, full rationality dictates that people must "anti-imitate" some of those they observe; and (ii) in every observation structure full rationality dictates that people who do not anti-imitate can, in essence, imitate at most one person among predecessors who share common information. We also show that in a very broad class of settings, virtually any learning rule in which people regularly do imitate more than one person without anti-imitating others will lead to a positive (and, in some environments, arbitrarily high) probability of people converging to confident and wrong long-run beliefs. When testing either the rationality or the efficiency of social learning, researchers should not focus on whether people follow others' behavior-but instead whether they follow it too much. (JEL B49) * We thank Paul Heidhues, seminar participants at Berkeley, DIW Berlin, Hebrew University, ITAM, LSE, Oxford, Penn, Tel Aviv University, and WZB, as well as the editor and referees, for their comments. We are grateful to the Russell Sage Foundation for its hospitality, and Katie Winograd and especially Claire Gabriel for their help with the historical literature on treating syphilis and other medical practices, and Tristan Gagnon-Bartsch and Min Zhang for excellent research assistance.